Architecture, Philosophy & Security Guarantees
FlowGuard AI was built from first principles for TestSprite Hackathon Season 4. Here is how our multi-agent architecture and verification engine work under the hood.
Core Principle: Evidence Precedes Verdicts
In traditional static AI reviewers, language models often output confident claims like "Vulnerability Verified" or "Code is Secure" based solely on token probability. In mission-critical software, this is dangerous hallucination.
FlowGuard enforces an uncompromising rule: an AI model can only suggest a POTENTIAL finding. A finding can only be designated as VERIFIED_PASS or VERIFIED_FAIL when the target state machine has been executed in our test harness under real concurrent conditions and the side effects have been physically counted.
Multi-Agent State Machine Pipeline
The lifecycle of an analysis run follows an explicit, audited sequence of transitions:
Agent A: Requirement Analyst
Accepts human business specifications and translates them into machine-verifiable mathematical invariants (e.g. count(side_effects) ≤ 1).
Agent B: Risk Hunter
Inspects source diffs specifically searching for race conditions, non-atomic database checks, and dirty read windows that could violate the extracted invariant.
Agent C: Verification Planner
Synthesizes concrete input vectors and runnable test specifications targeting discovered edge-cases, ready for execution in TestSprite 3.0.
Security Boundaries & Untrusted Input Isolation
No Arbitrary Shell Execution
External GitHub repositories are treated as completely untrusted input. FlowGuard NEVER executes arbitrary repository code or shell commands on the application server. Tests are run solely against first-party allowlisted synthetic fixtures.
Read-Only GitHub REST Access
Pull requests are inspected strictly via read-only GitHub REST API calls without write privileges. Input sizes are bounded to 100 KB diff and 20 files to prevent denial-of-service and memory exhaustion.
TestSprite Hackathon Season 4 Verification
FlowGuard AI connects directly to TestSprite 3.0 GitHub App to run automated PR checks and regression validation suites automatically.